Our chatting assist has drafted replies since 2022. It knows the fan's history, the model's voice, the boundaries she approved. It makes a good chatter faster. It has never once been allowed to press send, and that is the least interesting rule on its list.
Where the line actually is
The obvious rules came first: no promises the model hasn't approved, no prices invented on the spot, no pretending to be "really her" when a fan asks directly. A fan who asks is not testing the tech. He is asking whether the relationship is honest on its own terms. That answer is scripted by the model herself, in advance, in her words.
A fan can forgive a slow reply. He does not forgive a reply that could have been sent to anyone.
The list that matters
The messages we never automate are the expensive ones. Grief, a lost job, a birthday he mentioned once in March. The assist is built to detect those moments and do the one thing software should do there: stop, and hand the thread to the person on shift with everything they need to answer like they remember. Because they do — the note was kept by a human the first time, too.
What automation is for: recall, timing, arithmetic. What it is never for: the sentence a fan will read twice.
Why we publish this
Every few months an agency pitches a fully automated floor: no chatters, pure margin. We have tested every generation of that idea against our own numbers. Retention drops within six weeks, and it drops first among the top spenders, the fans who notice. The floor stays human because the numbers say so. The AI stays, too — one draft behind.